DeepSeek’s V4-Pro is officially out of preview. A changelog entry dated August 13, 2026 on DeepSeek’s API docs declares the GA release of DeepSeek-V4-Pro is live across the app, web, and API. The model name you call is still deepseek-v4-pro, the alias did not change, and the dated checkpoint underneath it is DeepSeek-V4-Pro-0813. MIT-licensed weights for that checkpoint landed the same day on Hugging Face at deepseek-ai/DeepSeek-V4-Pro-0813.
There is a lot to unpack, and most of it lands today.
The V4-Pro pricing cliff goes live at 16:00 UTC today
DeepSeek is shifting from flat API rates to a peak/off-peak model on the V4 family. The changelog says the change takes effect at 16:00 (UTC Time) on August 16, 2026. Off-peak rates are set at exactly half of peak-hour prices, and both tiers sit above today’s flat rates.
Per HyperAI’s coverage, V4-Pro at peak will run $1.32 per million input tokens and $3.96 per million output tokens. Off-peak is half of that. V4-Flash follows the same structure. The new pricing is the first time-of-day tiered pricing on a frontier Chinese model API, and it is meant to push non-urgent workloads into lower-demand windows.
DeepSeek has been near or at capacity since the V4 preview shipped. Tiered pricing is a load-shaping tool, not a goodwill gesture. If you route traffic through V4-Pro today, the bill you pay next month will look different depending on which hours your inference actually ran.
What GA actually changes
The model itself is what preview was promising, plus a post-training pass. The Hugging Face model card calls out DeepSeek-V4-Pro-0813 as “the official release of DeepSeek-V4-Pro, superseding the preview version, with greatly enhanced agentic capabilities and performance improvements that are especially pronounced in production environments.” It is built on the preview model structure with a DSpark speculative decoding module attached.
There are three concrete additions beyond the post-training bump.
First, native support for the OpenAI Responses API format, specifically adapted for Codex, with a one-click setup script. Any tool that already speaks Responses can target DeepSeek’s API without glue code.
Second, three thinking effort levels for both V4-Pro and V4-Flash: low, high, and max. DeepSeek says low for simple tasks, high for daily agent tasks, and max for more complex scenarios. Neither surface states which rung is the default, so if your costs depend on it, set the level explicitly rather than relying on the unset behavior.
Third, the weight release under MIT. The Hugging Face repo was created the same day as the changelog entry, and it ships a 1.7T-parameter checkpoint with the model card, encoding scripts, and vLLM/SGLang run recipes in one place. If you want to self-host V4-Pro today, you can.
Harness v0.1 is the agent piece
DeepSeek is also shipping a developer preview of DeepSeek Harness v0.1, an agentic coding scaffold that reads files, edits code, browses the web, and runs tasks to completion. That is the same layer Anthropic sells as Claude Code, and the same layer Cursor and OpenAI Codex sit on. The interesting detail is the framing: Harness has an open architecture where you can plug in any model, including rivals’ models, rather than hard-coding DeepSeek’s own. DeepSeek explicitly positioned this against “American rivals” that hard-code their products.
This is a direct move into the agentic coding market. The harness is in developer preview rather than GA, and pricing for Harness-mediated runs has not been published separately. Treat it as a capability preview, not a production lane yet.
Where the benchmarks land
The vendor benchmark table on the model card puts DeepSeek-V4-Pro-0813 close to frontier closed models on agent-heavy suites. Highlights include Terminal Bench 2.1 at 87.9 (vs Fable 5 at 88.0 and Opus 4.8 at 85.0), HLE with tools at 60.0 (vs Fable 5 at 63.0), Cybergym at 83.3 (vs Fable 5 at 83.1), and DSBench-Hard at 67.2. Those numbers are vendor-run on DeepSeek’s own harness with no independent reproduction at the time of writing. Hold them loosely until third-party evals land.
The structural detail worth more than a benchmark score is that the alias did not change. Anyone calling deepseek-v4-pro was moved onto the GA build without touching a line of code, and without necessarily knowing it happened. If your compliance process assumes the model behind an alias never changes without an announcement, this release is your counterexample.
What to do today
If you build on the DeepSeek API, three things matter today.
Map your current and planned inference loads to the new peak and off-peak windows before 16:00 UTC. Off-peak is half price, and your crons or orchestration rules should land expensive batch jobs there if your latency budget allows.
If you self-host, the Hugging Face weights are MIT-licensed and live now. Pull them before you commit, and check the DSpark speculative decoding recipe if you want inference speed on existing hardware.
If you build agent coding tools, DeepSeek Harness v0.1 is the layer worth tracking. Open-architecture agent scaffolds are rare from a frontier lab, and if the developer preview holds up, it will pull workload off the closed agent stacks over the next two quarters.
The pricing cliff is the story most teams will feel first. The GA build and the weights are the story most teams will remember.
